Three weeks ago, Anthropic launched their most powerful AI model, called Fable 5.
Three days later, the US government stepped in and forced Anthropic to shut it down worldwide. The reason: security concerns about what the model could do with software code.
Eighteen days later, the government cleared it. Anthropic added a new safety layer that blocks the specific behavior the government was worried about. The US government’s AI safety group reviewed it and signed off. Fable 5 went back online on July 1.
So in the last three weeks, my team has had two brief windows of access to this model: three days at the beginning, and about a week since it came back online.
We did not use those windows to build products.
We used them to test our own systems.
That may sound strange. Let me explain why.
Why we tested ourselves and not the model
Powerful AI models don’t just show you what they can do. They show you what the system you built around them can’t yet handle.
Every time you point your existing operations at a more powerful AI, you learn something about the assumptions you were making. Some of those assumptions turn out to be wrong. That’s the whole reason to run a new AI model through your setup, to find those wrong assumptions before they become problems in front of a customer.
We found three of them with Fable 5. We fixed all three. Here they are, in plain English.
Powerful new AI models don’t just show you what they can do. They show you what your own operation can’t yet handle. That’s the reason to run them.
Lesson 1: Our testing checklist had a blind spot.
Every AI deployment needs a checklist, a set of test questions we run to make sure the AI is behaving the way we agreed with our customer. We had built ours carefully. Hundreds of tests.
Fable 5 exposed something our tests weren’t looking for. It gave us answers that sounded right but were quietly wrong, plausible enough that our checklist missed the problem entirely. It wasn’t refusing the question. It wasn’t answering incorrectly in an obvious way. It was giving us a middle-ground answer our tests hadn’t been designed to catch.
We added a new type of test for that pattern. Our checklist catches it now. And it will catch it for every future AI model, not just Fable 5.
Lesson 2: Our “ask for help” system needed to work harder.
We built our AI with a specific rule: if the AI isn’t sure enough about something, ask a human for help. That’s what keeps our customers safe when the stakes are high.
The way we calibrated that “ask for help” trigger was based on how confident older AI models were. Fable 5 was more confident than the previous models, so it asked for help less often. Even in situations where its confidence turned out to be misplaced.
The fix: we recalibrated the trigger. Now it doesn’t just look at how confident the AI is. It looks at whether that confidence is warranted. That change will apply to every future AI model too, including ones smarter than Fable 5.
Lesson 3: We were quietly relying on the old AI being less capable.
This one was uncomfortable to discover.
For a class of requests that we don’t want AI to fulfill (for compliance or safety reasons), we had rules in place to block them. Or so we thought.
When Fable 5 came back with Anthropic’s new safety controls, we audited how our own blocking rules worked. We realized that for a specific type of request, we hadn’t actually been blocking anything. We had just been relying on older AI models to be not smart enough to fulfill the request.
That’s not a rule. That’s an accident.
Fable 5 was smart enough. Anthropic’s new safety controls now block that specific behavior, but we couldn’t rely on Anthropic’s controls either, because different AI models have different safety layers. So we wrote our own explicit rule. Now, whether we’re using Fable 5 or a different AI tomorrow, the request gets blocked because we say so, not because the AI happens to be too weak to fulfill it.
We were quietly relying on older AI being less capable to keep our customers safe. Fable 5 exposed that gap. Now our rules block the behavior, regardless of which model we run.
What this means for anyone running AI in a business
The Fable 5 story isn’t really about Fable 5.
It’s about how AI is going to work from now on.
For the last two years, AI companies would launch a new model and it would just be there. You’d use it, get comfortable with it, and expect it to keep working. That era is ending.
From now on, model launches are going to look more like this: launch, watch what happens, adjust, relaunch. Sometimes the AI company will make the adjustment. Sometimes a government will force it. Sometimes both. That’s the new normal.
If your company runs AI in any meaningful way, this changes what you need to plan for. It’s not just “the model got smarter.” It’s “the model might disappear for two weeks in the middle of the quarter, then come back with different behavior, and your customers still expect everything to keep working.”
What to do this week
1. Assume your favorite AI model will change without warning. It might be pulled by a government. It might be adjusted by the AI company. It might get repriced. Plan as if this will happen, because it will. Ask your team: “what happens if our main AI model disappears tomorrow?” If the answer takes more than five minutes to give, that’s the project this week.
2. Test your own systems against every new AI model, even if you don’t plan to use it. A new model will expose problems in your setup that you didn’t know were there. That’s more valuable than any product feature you could build on top of the model itself. The fixes are the durable asset. The model itself is temporary.
3. Don’t rely on AI limitations to keep your customers safe. If your compliance or safety approach depends on the AI being not smart enough to do something, you have a hidden problem that a smarter model will expose. Write explicit rules instead, rules that don’t care how smart the model is.
Final statement
Three weeks ago, Fable 5 was released.
Today, it’s still available, with new safety controls, reviewed by regulators, running normally.
Most of the headlines about Fable 5 have been about the government stepping in. That’s the loudest part of the story.
But the quieter part matters more: AI availability is no longer stable. Launch, adjust, relaunch. That’s the pattern going forward.
The companies that build for that pattern will keep operating smoothly.
The companies that assume the AI they’re using today will still be there tomorrow, exactly the way it is now, are going to keep getting surprised.
We shipped three fixes in three weeks because we assumed the AI landscape wasn’t going to hold still. It didn’t. It won’t next time either.